The Census Bureau’s Unorthodox Foray Into Voter Roll Investigations
Last week, President Trump used the official White House social media account to once again claim that he’d won the 2020 election (“I WON THE ELECTION!”). This time, his declaration stemmed from a seven-page Census Bureau report alleging at least 24,000 noncitizens voted in the 2020 general election (in which 154.6 million Americans voted).
At the risk of indulging the absurd: Trump’s claim isn’t supported. Even if we accept the report’s findings as facts (more on that below), it still doesn’t change the outcome of the 2020 presidential election. Of the states with the closest margins of victory for Joe Biden: Arizona (10,457 votes), Georgia (11,779), Wisconsin (20,682), Nevada (33,596), the Census report only alleges 1,100 (AZ), 400 (GA), 500 (WI), and 250 (NV) alleged noncitizens. And of course, we know that not every alleged unlawful noncitizen vote went for Biden.
But the real headline should be the extraordinary nature of the Census report (titled “Noncitizen Voting in the 2020 Election, A Beginning Analysis”). Never before has the Census Bureau studied the eligibility of individual American voters in a specific election. Never before has the Census Bureau investigated alleged voting crimes. Never before has the Census Bureau been deployed in furtherance of a president’s pet political project. And never before has the Census Bureau colored so far outside the lines of its popularly understood mandate (to collect broad, aggregate information about the American people and economy).
Even the publishing format is deviant: it’s an unsigned, incomplete report. Misty Heggeness, a former Census Bureau employee, told NPR that, “My biggest concern would be there’s no author attached to the report, which is not common practice within the Census Bureau.” Barry Burden, of the University of Wisconsin-Madison, said to the Associated Press that, “[c]ontrary to other Census publications, this report is unsigned, is described merely as a ‘beginning’ analysis, and relies on data from an unfamiliar private company.”
That nobody is willing to sign the report likely indicates that the report is a professional embarrassment or, as suggested by NPR’s reporting, that it’s not the work of Census Bureau career professionals and might even involve labor from the Trump-aligned think tank America First Policy Institute.
Regarding the report’s “incomplete” status, the Census Bureau gave no explanation for why it felt the need to publish a preliminary report before completing its analysis. Again, doing so is highly unusual. It could be an excuse for including sensational-but-half-baked theories that can later be excused (“well, we weren’t done with our analysis yet!”).
Then there’s the problematic methodology, at least as it is very briefly described on page six of the report. A number of smart people have already pointed out its numerous flaws, and I’ve quoted those at the bottom of this post. But as a table-setting matter, I’d caution against any large election data matching effort that works from a commercial source instead of the official county file, and I’d also caution against any matching effort that isn’t working from consistent, unique identifiers (e.g., using full name, date of birth, and social security number, or full name, date of birth, and driver’s license).
The Census Bureau report commits both sins. I oversaw the response of Maricopa County (Arizona) in 2021 to a similar effort performed by the technology firm Cyber Ninjas. The Cyber Ninjas used commercial data and didn’t use consistent, unique “hard-match” data. They alleged 34,278 problematic registration records. We went through each one. We agreed on only 37.
Previous attempts at large election data analysis should have humbled the federal government. False positives within the DHS SAVE program are well documented. A February 2026 investigation by ProPublica and The Texas Tribune begins: “It later turned out more than half the Boone County voters identified as noncitizens [by the SAVE program] were actually citizens.” Just last month, DHS Secretary Mullin sent letters to California, Nevada, New Jersey, and Pennsylvania alleging hundreds of thousands of noncitizens on their voter rolls. MS NOW reported this week that “more than a month later, the Department of Homeland Security is struggling to provide any proof.”
Such escapades prompted Utah’s (Republican) lieutenant governor, Deidre Henderson, to say, “These lists, and all of this hyperbole, are just meant to create a sense of urgency and chaos. I don’t know what they’re trying to do except make people feel afraid and make people distrust election results.”
Recent citizenship audits by states and counties that have painstakingly performed record-by-record investigations have found noncitizens registered to vote. But they make up a fraction of a fraction of a percent of the total voter rolls. For example, in May 2026, Utah announced the results of its audit of its 2.1 million voters. It found 27 noncitizens on the voter rolls, 13 of whom had voted in an election since 2018. In April 2025, Michigan identified 15 noncitizens who cast a ballot in the 2024 general election. Approximately 5.7 million people voted in Michigan’s November 2024 general election.
I continue to collect these state and local audits at Declare News.
It remains to be seen what the Census Bureau will do next on this front. Will it refer the allegedly unlawful voting records to the states? To law enforcement?
Stay tuned.
Other methodological commentary:
Justin Levitt (Election Law Blog): “There are three potential sources of error in that analysis: potential errors in the private vendor’s voter file, potential errors in the federal records, and potential errors in the matching process itself, known in the trade as record linkage. Real Census professionals are extremely careful and normally include caveats about potential sources of error in their work. The release didn’t mention any of these.”
Justin Levitt (Election Law Blog): “The complete absence of any mention of the error rates that swamp the actual findings: that’s the only red flag you need to know that this release isn’t worth anyone’s time.” Levitt (and others) note that the number of alleged noncitizens (24,000) is only 0.019% of the number of records reviewed by the report (128 million). Further, one of the sources cited by the Census Bureau report “expressly concerns false match rates in the record linkage process they used. And the vast majority of the errors that paper finds — in different contexts, comparing different sources — are larger than 0.019%.”
Justin Levitt (Election Law Blog): “There’s no note about why they stopped at 128 million records out of 160, rather than waiting for a full analysis of the whole, or what’s different about those records from the others. I could go on (ask me about the rounding protocol). The whole thing stinks.”
Gabriel Sterling (Declare News): “The analysis does not discuss derived citizens. Federal law, specifically 8 U.S.C.§1431, has been on the books for over a quarter century. It essentially allows for children of naturalized citizens to be citizens. There is not necessarily a record to confirm or check against; many derived citizens never file an N‑400 Form, nor take a naturalization oath, nor receive a Certificate of Citizenship.”
Gabriel Sterling (Declare News): “There are no false positive ratings given. With this much data, there are likely false positives. Whether that amount is large or small cannot be determined from this report.”
Mark Lindeman (Associated Press): “This is the weakest methodology description I’ve ever seen in a Census Bureau publication.”
Michael Morse (NPR): “I’d expect that the Census data matching [exercise] is plagued by the lack of unique identifiers in voter records and outdated or incomplete naturalization records. But the brief’s lack of transparency makes it difficult to pin down each problem or trust any explanation. Regardless, the estimated rate of apparent non-citizen voting is both a tiny fraction of total votes … and entirely consistent with both types of error.”